Development and Evaluation of an AI System for COVID-19

Cheng Jin; Weixiang Chen; Yukun Cao; Zhanwei Xu; Xin Zhang; Lei Deng; Chuansheng Zheng; Jie Zhou; Heshui Shi; Jianjiang Feng

huge dataset ImageNet7 for better and faster convergence. We tested a 3D classification network but this 2D scheme showed much better performance. The input of classification model is lung-masked slices, which means the input slices including training, internal validation and external test cohort, have been segment by segmentation model to get lung masks. The outputs of classification model are two scores respectively representing confidence levels of being normal and COVID-19 affected. Loss function of this block is cross entropy. The block was trained using 2D slices with batch size 32 for 100 epochs which costed about 5 hours under learning rate 10-5. Slices for training this block were extracted from training cohort, and the extraction process is detailed explained in supplementary methods. Since the predictions are based on 2D slices, an extra step is done to get a volume-level prediction. Because one volume is COVID-19 positive when any one of its slices is COVID-19 positive, we averaged the top 3 highest scores of all slices of a volume as the volume score. As a result, though training and validation were done on slice level, the block can take the whole CT volumes (with the whole lung segmentation volumes) and output a single prediction on volume level.